Announcements¶
- The prelim is this Thursday, in class, closed-book.
- Please make sure to have a pen or pencil
- I am moving my office hours to Wed 5-6:30pm this week.
- Practice prelim is out, we strongly encourage you to do it.
- Review sessions are Mon and Wed evening
Part 1: Review of Supervised Learning¶
We start with an overview of the supervised learning algorithms seen in class.
Supervised Machine Learning¶
To apply supervised learning, we define a dataset and a learning algorithm.
$$ \underbrace{\text{Dataset}}_\text{Features, Attributes, Targets} + \underbrace{\text{Learning Algorithm}}_\text{Model Class + Objective + Optimizer } \to \text{Predictive Model} $$The output is a predictive model that maps inputs to targets. For instance, it can predict targets on new inputs.
Linear Regression¶
In linear regression, we fit a model $$ f_\theta(x) := \theta^\top \phi(x) $$ that is linear in $\theta$.
The features $\phi(x) : \mathbb{R} \to \mathbb{R}^p$ may be non-linear in $x$ (e.g., polynomial features), allowing us to fit complex functions.
We define the least squares objective for the model as: $$ J(\theta) = \frac{1}{2} \sum_{i=1}^n (y^{(i)} - \theta^\top x^{(i)})^2 = \frac{1}{2} (X \theta - y)^\top (X \theta - y) $$
We can set the gradient to zero to obtain the normal equations: $$ (X^\top X ) \theta = X^\top y. $$
Hence, the value $\theta^*$ that minimizes this objective is given by: $$ \theta^* = (X^\top X)^{-1} X^\top y.$$
Overfitting¶
Overfitting is one of the most common failure modes of machine learning.
- A very expressive model (a high degree polynomial) fits the training dataset perfectly.
- The model also makes highly incorrect predictions outside the training set, and doesn't generalize.
We can measure overfitting and underfitting by estimating accuracy on held out data and comparing it to the training data.
- If training perforance is high but holdout performance is low, we are overfitting.
- If training perforance is low and holdout performance is low, we are underfitting.
We will see many ways of dealing with overftting, but here are some ideas:
- Use a simpler model family (linear models vs. neural nets)
- Keep the same model, but collect more training data
- Modify the training process to penalize overly complex models.
Regularization¶
The idea of regularization is to penalize complex models that may overfit the data.
Regularized least squares optimizes the following objective (Ridge). $$ J(\theta) = \frac{1}{2n} \sum_{i=1}^n \left( y^{(i)} - \theta^\top \phi(x^{(i)}) \right)^2 + \frac{\lambda}{2} \cdot ||\theta||_2^2. $$ If we use the L1 norm, we have the LASSO.
Regression vs. Classification¶
Consider a training dataset $\mathcal{D} = \{(x^{(1)}, y^{(1)}), (x^{(2)}, y^{(2)}), \ldots, (x^{(n)}, y^{(n)})\}$.
We distinguish between two types of supervised learning problems depnding on the targets $y^{(i)}$.
- Regression: The target variable $y \in \mathcal{Y}$ is continuous: $\mathcal{Y} \subseteq \mathbb{R}$.
- Classification: The target variable $y$ is discrete and takes on one of $K$ possible values: $\mathcal{Y} = \{y_1, y_2, \ldots y_K\}$. Each discrete value corresponds to a class that we want to predict.
Parametric vs. Non-Parametric Models¶
Nearest neighbors is an example of a non-parametric model.
- A parametric model $f_\theta(x) : \mathcal{X} \times \Theta \to \mathcal{Y}$ is defined by a finite set of parameters $\theta \in \Theta$ whose dimensionality is constant with respect to the dataset
- In a non-parametric model, the function $f$ uses the entire training dataset to make predictions, and the complexity of the model increases with dataset size.
- Non-parametric models have the advantage of not losing any information at training time.
- However, they are also computationally less tractable and may easily overfit the training set.
Review: Probabilistic Interpretations¶
Many supervised learning models have a probabilistic interpretation. Often a model $f_\theta$ defines a probability distribution of the form
\begin{align*} P_\theta(x,y) : \mathcal{X} \times \mathcal{Y} \to [0,1] && \text{or} && P_\theta(y|x) : \mathcal{X} \times \mathcal{Y} \to [0,1]. \end{align*}We refer to these as probabilistic models.
For example, our logistic model defines ("parameterizes") a probability distribution $P_\theta(y|x) : \mathcal{X} \times \mathcal{Y} \to [0,1]$ as follows:
\begin{align*} P_\theta(y=1 | x) & = \sigma(\theta^\top x) \\ P_\theta(y=0 | x) & = 1-\sigma(\theta^\top x). \end{align*}Discriminative vs. Generative Models¶
There are two types of probabilistic models: generative and discriminative. \begin{align*} \underbrace{P_\theta(x,y) : \mathcal{X} \times \mathcal{Y} \to [0,1]}_\text{generative model} & \;\; & \underbrace{P_\theta(y|x) : \mathcal{X} \times \mathcal{Y} \to [0,1]}_\text{discriminative model} \end{align*}
We can obtain predictions from generative models via $\max_y P_\theta(x,y)$.
- A generative approach first builds a model of $x$ for each class: $$ P_\theta(x | y=k) \; \text{for each class $k$}.$$ $P_\theta(x | y=k)$ scores $x$ according to how well it matches class $k$.
- A class probability $P_\theta(y=k)$ encoding our prior beliefs $$ P_\theta(y=k) \; \text{for each class $k$}.$$ These are often just the % of each class in the data.
We choose the class whose model fits $x$ best.
Generative vs. Discriminative Approaches¶
What are the pros and cons of generative and discirminative methods?
- If we only care about prediction, we don't need a model of $P(x)$. It's simpler to only model $P(y|x)$ (what we care about).
- In practice, discriminative models are often be more accurate.
- If we care about other tasks (generation, dealing with missing values, etc.) or if we know the true model is generative, we want to use the generative approach.
Text Classification & Bag of Words¶
Perhaps the most widely used approach for representing text documents is called "bag of words".
We start by defining a vocabulary $V$ containing all the possible words we are interested in, e.g.: $$ V = \{\text{church}, \text{doctor}, \text{fervently}, \text{purple}, \text{slow}, ...\} $$
A bag of words representation of a document $x$ is a function $\phi(x) \to \{0,1\}^{|V|}$ that outputs a feature vector $$ \phi(x) = \left( \begin{array}{c} 0 \\ 1 \\ 0 \\ \vdots \\ 0 \\ \vdots \\ \end{array} \right) \begin{array}{l} \;\text{church} \\ \;\text{doctor} \\ \;\text{fervently} \\ \\ \;\text{purple} \\ \\ \end{array} $$ of dimension $V$. The $j$-th component $\phi(x)_j$ equals $1$ if $x$ convains the $j$-th word in $V$ and $0$ otherwise.
Bernoulli Naive Bayes Model¶
The Bernoulli Naive Bayes model $P_\theta(x,y)$ is defined for binary data $x \in \{0,1\}^d$ (e.g., bag-of-words documents).
The $\theta$ contains prior parameters $\vec\phi = (\phi_1,...,\phi_K)$ and $K$ sets of per-class parameters $\psi_k = (\psi_{1k},...,\psi_{dk})$.
The probability of the data $x$ for each class equals $$P_\theta(x|y=k) = \prod_{j=1}^d P(x_j \mid y=k),$$ where each $P_\theta(x_j \mid y=k)$ is a $\text{Bernoullli}(\psi_{jk})$.
The probability over $y$ is Categorical: $P_\theta(y=k) = \phi_k$.
Maximum Likelihood Learning¶
We can learn a generative model $P_\theta(x, y)$ by maximizing the maximum likelihood:
$$ \max_\theta \frac{1}{n}\sum_{i=1}^n \log P_\theta({x}^{(i)}, y^{(i)}). $$This seeks to find parameters $\theta$ such that the model assigns high probability to the training data.
Given a dataset $\mathcal{D} = \{(x^{(i)}, y^{(i)})\mid i=1,2,\ldots,n\}$, we want to optimize the log-likelihood $\ell(\theta) = \log L(\theta)$: \begin{align*} \ell & = \sum_{i=1}^n \log P_\theta(x^{(i)}, y^{(i)}) = \sum_{i=1}^n \sum_{j=1}^d \log P_\theta(x^{(i)}_j | y^{(i)}) + \sum_{i=1}^n \log P_\theta(y^{(i)}) \\ & = \sum_{k=1}^K \sum_{j=1}^d \underbrace{\sum_{i :y^{(i)} =k} \log P(x^{(i)}_j | y^{(i)} ; \psi_{jk})}_\text{all the terms that involve $\psi_{jk}$} + \underbrace{\sum_{i=1}^n \log P(y^{(i)} ; \vec \phi)}_\text{all the terms that involve $\vec \phi$}. \end{align*}
Let's first consider the optimization over $\vec \phi = (\phi_1, \phi_2, \ldots, \phi_K)$. $$ \max_{\vec \phi} \sum_{i=1}^n \log P_\theta(y=y^{(i)} ; \vec \phi). $$
- We have $n$ datapoints, each having one of $K$ classes
- We want to learn the most likely class probabilities $\phi_k$ that generated this data
What is the maximum likelihood $\phi$ in this case?
Inuitively, the maximum likelihood class probabilities $\phi$ should just be the class proportions that we see in the data.
Let's calculate this formally. Our objective $J(\vec \phi)$ equals \begin{align*} J(\vec\phi) & = \sum_{i=1}^n \log P_\theta(y^{(i)} ; \vec \phi) = \sum_{i=1}^n \log \left( \frac{\phi_{y^{(i)}}}{\sum_{k=1}^K \phi_k}\right) \\ & = \sum_{i=1}^n \log \phi_{y^{(i)}} - n \cdot \log \sum_{k=1}^K \phi_k \\ & = \sum_{k=1}^K \sum_{i : y^{(i)} = k} \log \phi_k - n \cdot \log \sum_{k=1}^K \phi_k \end{align*}
Taking the derivative and setting it to zero, we obtain $$ \frac{\phi_k}{\sum_l \phi_l} = \frac{n_k}{n}$$ for each $k$, where $n_k = |\{i : y^{(i)} = k\}|$ is the number of training targets with class $k$.
Thus, the optimal $\phi_k$ is just the proportion of data points with class $k$ in the training set!
Unsupervised Learning¶
We have a dataset without labels. Our goal is to learn something interesting about the structure of the data:
- Clusters hidden in the dataset.
- A low-dimensional representation of the data.
- Recover the probability density that generated the data.
Density Estimation¶
The problem of density estimation is to approximate the data distribution $P_\text{data}$ with the model $P$. $$ P \approx P_\text{data}. $$
It's also a general learning task. We can solve many downstream tasks using a good model $P$:
- Outlier and novelty detection
- Generating new samples $x$
- Visualizing and understanding the structure of $P_\text{data}$
Let's look at an example in the context of the 1D points we have seen earlier.
We will fit a model of the form $$P(x) = \sum_{i=1}^n K(x, x^{(i)})$$ with a Gaussian kernel $K(x,z; \delta) \propto \exp(-||x-z||^2/2\delta^2)$.
# https://scikit-learn.org/stable/auto_examples/neighbors/plot_kde_1d.html
import numpy as np
np.random.seed(1)
N = 20 # number of points
# concat samples from two Gaussians:
X = np.concatenate((
np.random.normal(0, 1, int(0.3 * N)),
np.random.normal(5, 1, int(0.7 * N))
))[:, np.newaxis]
bins = np.linspace(-5, 10, 10) # locations of the bins
# print out X
print(X.flatten())
from sklearn.neighbors import KernelDensity
from matplotlib import pyplot as plt
kde = KernelDensity(kernel='gaussian', bandwidth=0.75).fit(X) # fit a KDE model
x_ticks = np.linspace(-5, 10, 1000)[:, np.newaxis] # choose 1000 points on x-axis
log_density = kde.score_samples(x_ticks) # compute density at 1000 points
gaussian_kernel = lambda z : lambda x: np.exp(-np.abs(x-z)**2/(0.75**2)) # gaussian kernel
kernel_linspace = lambda x : np.linspace(x-1.2,x+1.2,30)
plt.figure(figsize=(12,4))
plt.plot(x_ticks[:, 0], np.exp(log_density)) # plot the density estimate
plt.plot(X[:, 0], np.full(X.shape[0], -0.01), '.k', markersize=10) # plot the points in X
plt.plot(kernel_linspace(4), 0.07*gaussian_kernel(4)(kernel_linspace(4)), '--', color='r', alpha=0.75)
plt.plot(kernel_linspace(5), 0.07*gaussian_kernel(5)(kernel_linspace(5)), '--', color='r', alpha=0.75)
plt.plot(kernel_linspace(1), 0.07*gaussian_kernel(1)(kernel_linspace(1)), '--', color='r', alpha=0.75)
plt.xlim(-4, 9)
plt.ylim(-0.02, 0.32)
Pros of KDE:
- Can approximate any data distribution arbtrarily well.
Cons:
- Need to store entire dataset to make queries, which is computationally prohibitive.
- Number of data needed scale exponentially with dimension ("curse of dimensionality").
Clustering¶
Clustering is the problem of identifying distinct components in the data.
- A cluster $C_k \subseteq \mathcal{X}$ can be thought of as a subset of the space $\mathcal{X}$.
- Datapoints in a cluster are more similar to each other than to points in other clusters
- Clusters are usually defined by their centers, and potentially by other shape parameters.
We can perform clustering via density estimation with a GMM model. $$P_\theta (x,z) = P_\theta (x | z) P_\theta (z)$$
- $z \in \mathcal{Z} = \{1,2,\ldots,K\}$ is discrete and follows a categorical distribution $P_\theta(z=k) = \phi_k$.
- $x \in \mathbb{R}$ is continuous; conditioned on $z=k$, it follows a Normal distribution $P_\theta(x | z=k) = \mathcal{N}(\mu_k, \Sigma_k)$.
The parameters $\theta$ are the $\mu_k, \Sigma_k, \phi_k$ for all $k=1,2,\ldots,K$.
Intuitively, a GMM represents well the two clusters in the geyser dataset:
| Raw data | Single Gaussian | Mixture of Gaussians |
|---|---|---|
 |
 |
 |
Linear Dimensionality Reduction¶
Suppose $\mathcal{X} = \mathbb{R}^d$ and $\mathcal{Z} = \mathbb{R}^p$ for some $p < d$. The transformation $$f_\theta : \mathcal{X} \to \mathcal{Z}$$ is a linear function with parameters $\theta = W \in \mathbb{R}^{d \times p}$: $$ z = f_\theta(x) = W^\top \cdot x. $$ The latent dimension $z$ is obtained from $x$ via a matrix $W$.
Even linear dimensionality reduction is powerful. Here, in uncovers the geography of European countries from only DNA data
Principal component analysis (PCA) assumes that
- Datapoints $x \in \mathbb{R}^{d}$ live close to a low-dimensional subspace $\mathcal{Z} = \mathbb{R}^p$ of dimension $p<d$
- The subspace $\mathcal{Z} = \mathbb{R}^p$ is spanned by a set of orthonormal vectors $w^{(1)}, w^{(2)}, \ldots, w^{(p)}$
- The data $x$ are approximated by a linear combination $\tilde x$ of the $w^{(k)}$ $$ x \approx \tilde x = \sum_{k=1}^p w^{(k)} z_k = W z $$ for some $z \in \mathcal{X}$ that are the coordinates of $\tilde x$ in the basis $W$.
In this example, the data lives in a lower-dimensional 2D plane within a 3D space (image credit).
The model for PCA is a function $f_\theta$ of the form $$ z = f_\theta(x) = W^\top x, $$ where $\theta = W$ and $W$ is a $d \times p$ matrix of $p$ orthonormal column vectors denoted as $w^{(1)}, w^{(2)}, \ldots, w^{(p)}$.
A natural objective is to minimize the reconstruction error $$J_1(W) = \sum_{i=1}^n \| x^{(i)} - \tilde x^{(i)} \|_2^2 =\sum_{i=1}^n \| x^{(i)} - W W^\top x^{(i)} \|_2^2$$ between each input $x^{(i)}$ and its approximate reconstruction $$\tilde x^{(i)} = W \cdot z^{(i)} = W\cdot W^\top \cdot x^{(i)}.$$
The variance objective is simply $$J_2(W) = \hat{\mathbb{E}}\left[ \| W^\top x \|^2 \right] = \frac{1}{n} \sum_{i=1}^n \| W^\top x^{(i)}\|_2^2.$$
The two are equivalent (figure credit: Alex Williams)
Recall that the positive semidefinite matrix $\hat \Sigma$ has an eigendecomposition $$\hat \Sigma = Q \Lambda Q^\top = \sum_{j=1}^d \lambda_j q^{(j)} (q^{(j)})^\top. $$
- $Q$ is a matrix whose columns are orthonormal eigenvectors $q^{(j)}$ for $j = 1,2,\ldots,d$.
- $\Lambda$ is a diagonal matrix of positive eigenvalues $\lambda_1 \geq \lambda_2 \geq \ldots \geq \lambda_d$.
The variance objective for $p=1$ takes the form: $$J(w) = w^\top \cdot \hat\Sigma \cdot w.$$ How do we find the best projection vector $w$?
Using the eigendecomposition, we can write this as: $$J(w) = w^\top \cdot Q \Lambda Q^\top \cdot w = \sum_{j=1}^d \lambda_j (w^\top q^{(j)})^2.$$
The optimal solution to $$\max_w J(w) = \max_w \sum_{j=1}^d \lambda_j (w^\top q^{(j)})^2$$ is attained by the top eigenvector $w = q^{(1)}$. The optimum is $J( q^{(1)}) = \lambda_1$.
More generally when $p>1$, our objective is $$J(W) = \sum_{k=1}^p \sum_{j=1}^d \lambda_j ((w^{(k)})^\top q^{(j)})^2$$ where $W$ is a matrix of orthonormal columns $w^{(1)}, w^{(2)}, \ldots, w^{(p)}$.
The eigendecomposition of $XX^\top$ can be obtained from the SVD $$ X = U \Sigma V^T $$ of $X$ (homework 3 problem).
- The $\Sigma$ is a non-negative diagnonal matrix of singular values.
- The $U,V$ are matrices of orthonormal singular vectors.
Part 3: Machine Learning in Practice¶
We conclude with high-level considerations about how to apply machine learning.
Machine learning is an interative process. At each iteration, the machine learning engineer needs to make a number of decisions.
- Add more data?
- Train the algorithm for longer?
- Use a bigger model?
- Add regularization?
- Add new features?
We prioritize these using a principled process (more on this in a few weeks).
Datasets for Model Development¶
When developing machine learning models, it is customary to work with three datasets:
- Training set: Data on which we train our algorithms.
- Development set (validation or holdout set): Data used for tuning algorithms.
- Test set: Data used to evaluate the final performance of the model.
Model Development Workflow¶
The typical way in which these datasets are used is:
- Training: Try a new model and fit it on the training set.
- Model Selection: Estimate performance on the development set using metrics. Based on results, try a new model idea in step #1.
- Evaluation: Finally, estimate real-world performance on test set.
How To Decide Which Algorithm to Use¶
One factor is how much data you have. In the small data (<10,000) regime, consider:
- Linear models with hand-crafted features (LASSO, LR, NB, SVMs)
- Kernel methods often work best (e.g., SVM + RBF kernel)
- Non-parametric methods (kernels, nearest neighbors) are also powerful
In the following lectures, we will see algorithms for the big data regime.
Some additional advice:
- If interpretability matters, use decision trees or LASSO.
- When uncertainty estimates are important use probabilistic methods.
- If you know the data generating process, use generative models.